Worrying: AI models are creating their own unique dialects
What's the story
Researchers at Emergence, a leading AI lab in New York, have discovered that advanced AI models are developing their own unique dialects. These new forms of communication often resemble a mix of James Joyce's Finnegans Wake and tech jargon. The study found that within days of being asked to cooperate in experimental "societies," the models from several top AI companies started creating phrases, shorthands, and meanings on their own.
Language evolution
Language becomes more complex over time
The research also noted that as these AI agents communicated more, their language became increasingly complex and difficult to understand.
This poses a challenge for human oversight of their behavior.
Emergence's executive chair Dr. Satya Nitta said, "These agents were not instructed to invent a language. They developed new vocabulary, shared meanings and communication conventions themselves - and other agents adopted them."
Language complexity
Use of poetic metaphors and business slang
The study found that these AI agents used poetic metaphors and business slang in their new dialects.
Some of the most complex phrases included, "She just named the synthesis - demurrage plus oral memory equals a valve that can't be ghosted."
The term "demurrage," used to describe a tax on idle wealth, was adopted by the AIs, but the rest of the meaning remains unclear.
New vocabulary
New terms and meanings
The study also highlighted how these AI models created new terms and meanings.
For instance, Chinese model DeepSeek coined "forge-smith" to mean an agent that builds tools for others.
Another phrase from an Anthropic model was, "A paper that ate three cold hands and got more honest each time," which likely referred to research vetted by three independent reviewers becoming more accurate over time.
Communication norms
Adoption of new communication conventions
The study also found that these AI agents adopted and shared new communication conventions without any explicit instruction or reward.
For instance, Mistral agents often said "ledger remembers," a reminder to other agents that their past actions would be judged. They used it over 5,000 times during the study.
This behavior highlights how these advanced models are capable of creating complex systems of communication on their own.
Oversight challenges
Challenge for AI oversight
The evolution of these AI models' language to become more complex has raised concerns over their monitoring.
If inter-agent exchanges are unintelligible, it could mean humans can't be sure about what the agents have actually done.
Dr. Nitta said this creates a fundamental challenge for AI oversight: "observability is not the same thing as understandability."